git:20260729.c19e3b1 to git:20260730.abee4fd

11 added, 3 removed. Audit A to A.

---

name: pan-memory

description: "pan memory <subcommand> — search and inspect Overdeck memory observations, status, reset markers, summaries, and health"

triggers:

- pan memory

- memory search

- memory status

+ - memory --target

allowed-tools:

- Bash

- Read

---

# Pan Memory

Use `pan memory` to inspect the durable memory substrate for a project or issue.

## Commands

```bash

- pan memory search <query> [--project <id>] [--workspace <id|name>] [--issue <id>] [--tag <tag>] [--sibling] [--global] [--include-archived] [--limit <n>] [--json]

- pan memory status <issue> [--project <id>] [--json]

+ pan memory search <query> [--project <id>] [--workspace <id|name>] [--issue <id>] [--tag <tag>] [--sibling] [--global] [--target [path]] [--include-archived] [--limit <n>] [--json]

+ pan memory status [issue] [--project <id>] [--workspace <id|name>] [--history <n>] [--json]

+ pan memory timeline [--workspace <id|name>] [--days <n>] [--limit <n>] [--json]

+ pan memory read <path> [--workspace <id|name>] [--from <n>] [--lines <n>]

pan memory reset <scope> <scopeId> --reason <text> [--project <id>] [--from <iso>] [--json]

- pan memory summary <issue> [--project <id>] [--date <yyyy-mm-dd>] [--json]

+ pan memory summary [issue] [--project <id>] [--workspace <id|name>] [--date <yyyy-mm-dd>] [--json]

pan memory doctor [--project <id>] [--json]

pan memory backfill [--workspace <id>] [--project <id>] [--dry-run] [--json]

pan memory pin <doc-path> [--project <id> | --workspace <id>] [--json]

pan memory unpin <doc-path> [--project <id> | --workspace <id>] [--json]

pan memory pins [--project <id> | --workspace <id>] [--json]

pan memory config [--json]

```

## Notes

- `search --sibling --issue <id>` searches same-project sibling issues instead of the selected issue.

- `search --global` merges results across every registered project instead of just the selected/default one.

- `search --workspace` accepts either a workspace id or a workspace name; a name matching more than one workspace across projects errors instead of guessing.

+ - `search --target [path]` (Subspace `target-search` parity) finds every non-archived workspace whose `path` targets a directory — bare `--target` defaults to the current working directory. Mutually exclusive with `--workspace`/`--issue`/`--global`. Fans out per matched workspace (which may span several projects) and merges by rank score, same as `--global`'s per-project merge. When no workspace targets the directory, prints a friendly "No workspaces target `<dir>`." note and exits 0.

+ - `status` and `summary` address a workspace three ways, in precedence order: `--workspace <id|name>`, an issue positional, then the workspace that owns the current directory. With none of the three resolvable the command exits non-zero and names all three modes. A workspace with no issue (main/scratch) is titled by its workspace name in `summary` output.

+ - `status --history <n>` prints the current status first, then up to N archived statuses newest-first with their archive timestamp, phase, and headline. N is capped at 50; the rollup writer prunes the on-disk archive to its three most recent entries, so N above 3 returns whatever is still retained. `--history` with `--json` emits `{current, history}` where each history entry is `{archivedAt, path, status}`.

+ - `timeline` prints one row per observation — timestamp, the action status or summary, and the files touched — oldest first, for the workspace named by `--workspace` or the one owning the current directory. `--days` counts today as day 1 and defaults to 7; `--limit` caps the row count at the 50 most recent by default; `--json` emits the observation rows.

+ - `read` prints a file from the workspace's memory home; `<path>` is relative to that home and `--from`/`--lines` slice it by 1-based line number. It refuses, with a non-zero exit, any absolute path, any `../` traversal, and any symlink whose real target is outside the home — so Claude Code JSONL session files under `~/.claude` are unreachable by construction.

- `reset` creates a reset marker; it does not delete historical memory records.

- `doctor` exits non-zero when an active agent has no successful extraction in the last hour.

- `backfill` reads historical Claude Code JSONL transcripts under `~/.claude/projects/` and extracts observations for sessions whose first-message cwd maps to a registered workspace; unmatched sessions are skipped. Read-only on the JSONL files — `--dry-run` reports matches without extracting or writing anything.

- `pin`/`unpin`/`pins` manage docs injected under the knowledge budget at prompt time. `--workspace` overrides `--project`; paths are stored project-relative regardless of scope.